[AMDGPU] Handle f16 select{_cc}

- Select `select` to `v_cndmask_b32`
- Expand `select_cc`
- Refactor patterns

Differential Revision: https://reviews.llvm.org/D26714

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@287074 91177308-0d34-0410-b5e6-96231b3b80d8

diff --git a/lib/Target/AMDGPU/SIISelLowering.cpp b/lib/Target/AMDGPU/SIISelLowering.cpp
index a0650d407ef1576df3d42417826574012d0fe448..65898fc4b2a8740da29c7a2088c7bd2a13522139 100644 (file)
--- a/lib/Target/AMDGPU/SIISelLowering.cpp
+++ b/lib/Target/AMDGPU/SIISelLowering.cpp
@@ -289,6 +289,7 @@ SITargetLowering::SITargetLowering(const TargetMachine &TM,
     setOperationAction(ISD::FSIN, MVT::f16, Promote);
 
     // F16 - VOP2 Actions.
+    setOperationAction(ISD::SELECT_CC, MVT::f16, Expand);
     setOperationAction(ISD::FMAXNUM, MVT::f16, Legal);
     setOperationAction(ISD::FMINNUM, MVT::f16, Legal);
     setOperationAction(ISD::FDIV, MVT::f16, Promote);

diff --git a/lib/Target/AMDGPU/SIInstructions.td b/lib/Target/AMDGPU/SIInstructions.td
index 423599d8ccf14448211e5305e4cd37c830729fb1..14e4b73731583b8a58745e2b8a5bd4bb63788448 100644 (file)
--- a/lib/Target/AMDGPU/SIInstructions.td
+++ b/lib/Target/AMDGPU/SIInstructions.td
@@ -468,16 +468,23 @@ def : Pat <
 // VOP2 Patterns
 //===----------------------------------------------------------------------===//
 
+multiclass SelectPat<ValueType vt, Instruction inst> {
+  def : Pat <
+    (vt (select i1:$src0, vt:$src1, vt:$src2)),
+    (inst $src2, $src1, $src0)
+  >;
+}
+
+defm : SelectPat<i16, V_CNDMASK_B32_e64>;
+defm : SelectPat<i32, V_CNDMASK_B32_e64>;
+defm : SelectPat<f16, V_CNDMASK_B32_e64>;
+defm : SelectPat<f32, V_CNDMASK_B32_e64>;
+
 def : Pat <
   (i32 (add (i32 (ctpop i32:$popcnt)), i32:$val)),
   (V_BCNT_U32_B32_e64 $popcnt, $val)
 >;
 
-def : Pat <
-  (i32 (select i1:$src0, i32:$src1, i32:$src2)),
-  (V_CNDMASK_B32_e64 $src2, $src1, $src0)
->;
-
 // Pattern for V_MAC_F16
 def : Pat <
   (f16 (fmad (VOP3NoMods0 f16:$src0, i32:$src0_modifiers, i1:$clamp, i32:$omod),
@@ -990,11 +997,6 @@ def : Pat <
              (V_ALIGNBIT_B32 $a, $a, (i32 8)))
 >;
 
-def : Pat <
-  (f32 (select i1:$src2, f32:$src1, f32:$src0)),
-  (V_CNDMASK_B32_e64 $src0, $src1, $src2)
->;
-
 multiclass BFMPatterns <ValueType vt, InstSI BFM, InstSI MOV> {
   def : Pat <
     (vt (shl (vt (add (vt (shl 1, vt:$a)), -1)), vt:$b)),

diff --git a/lib/Target/AMDGPU/VOP3Instructions.td b/lib/Target/AMDGPU/VOP3Instructions.td
index 4f8b680e52ff73e3ea6eeac2adf72dc43ff3d7c3..5efa64d25ce1543e4a1a25d009862dd259c5ca7a 100644 (file)
--- a/lib/Target/AMDGPU/VOP3Instructions.td
+++ b/lib/Target/AMDGPU/VOP3Instructions.td
@@ -232,11 +232,6 @@ def V_MAD_I16 : VOP3Inst <"v_mad_i16", VOP3_Profile<VOP_I16_I16_I16_I16>>;
 
 } // End SubtargetPredicate = isVI
 
-def : Pat <
-  (i16 (select i1:$src0, i16:$src1, i16:$src2)),
-  (V_CNDMASK_B32_e64 $src2, $src1, $src0)
->;
-
 let Predicates = [isVI] in {
 
 multiclass Tenary_i16_Pats <SDPatternOperator op1, SDPatternOperator op2,

diff --git a/test/CodeGen/AMDGPU/select.f16.ll b/test/CodeGen/AMDGPU/select.f16.ll
new file mode 100644 (file)
index 0000000..a0ad475
--- /dev/null
+++ b/test/CodeGen/AMDGPU/select.f16.ll
@@ -0,0 +1,322 @@
+; RUN: llc -march=amdgcn -verify-machineinstrs < %s | FileCheck -check-prefix=GCN -check-prefix=SI %s
+; RUN: llc -march=amdgcn -mcpu=fiji -verify-machineinstrs < %s | FileCheck -check-prefix=GCN -check-prefix=VI %s
+
+; GCN-LABEL: {{^}}select_f16
+; GCN: buffer_load_ushort v[[A_F16:[0-9]+]]
+; GCN: buffer_load_ushort v[[B_F16:[0-9]+]]
+; GCN: buffer_load_ushort v[[C_F16:[0-9]+]]
+; GCN: buffer_load_ushort v[[D_F16:[0-9]+]]
+; SI:  v_cvt_f32_f16_e32 v[[A_F32:[0-9]+]], v[[A_F16]]
+; SI:  v_cvt_f32_f16_e32 v[[B_F32:[0-9]+]], v[[B_F16]]
+; SI:  v_cmp_lt_f32_e32 vcc, v[[A_F32]], v[[B_F32]]
+; SI:  v_cvt_f32_f16_e32 v[[C_F32:[0-9]+]], v[[C_F16]]
+; SI:  v_cvt_f32_f16_e32 v[[D_F32:[0-9]+]], v[[D_F16]]
+; SI:  v_cndmask_b32_e32 v[[R_F32:[0-9]+]], v[[D_F32]], v[[C_F32]]
+; SI:  v_cvt_f16_f32_e32 v[[R_F16:[0-9]+]], v[[R_F32]]
+; VI:  v_cmp_lt_f16_e32 vcc, v[[A_F16]], v[[B_F16]]
+; VI:  v_cndmask_b32_e32 v[[R_F16:[0-9]+]], v[[D_F16]], v[[C_F16]], vcc
+; GCN: buffer_store_short v[[R_F16]]
+; GCN: s_endpgm
+define void @select_f16(
+    half addrspace(1)* %r,
+    half addrspace(1)* %a,
+    half addrspace(1)* %b,
+    half addrspace(1)* %c,
+    half addrspace(1)* %d) {
+entry:
+  %a.val = load half, half addrspace(1)* %a
+  %b.val = load half, half addrspace(1)* %b
+  %c.val = load half, half addrspace(1)* %c
+  %d.val = load half, half addrspace(1)* %d
+  %fcmp = fcmp olt half %a.val, %b.val
+  %r.val = select i1 %fcmp, half %c.val, half %d.val
+  store half %r.val, half addrspace(1)* %r
+  ret void
+}
+
+; GCN-LABEL: {{^}}select_f16_imm_a
+; GCN: buffer_load_ushort v[[B_F16:[0-9]+]]
+; GCN: buffer_load_ushort v[[C_F16:[0-9]+]]
+; GCN: buffer_load_ushort v[[D_F16:[0-9]+]]
+; SI:  v_cvt_f32_f16_e32 v[[A_F32:[0-9]+]], 0x3800{{$}}
+; SI:  v_cvt_f32_f16_e32 v[[B_F32:[0-9]+]], v[[B_F16]]
+; SI:  v_cmp_gt_f32_e32 vcc, v[[B_F32]], v[[A_F32]]
+; SI:  v_cvt_f32_f16_e32 v[[C_F32:[0-9]+]], v[[C_F16]]
+; SI:  v_cvt_f32_f16_e32 v[[D_F32:[0-9]+]], v[[D_F16]]
+; SI:  v_cndmask_b32_e32 v[[R_F32:[0-9]+]], v[[D_F32]], v[[C_F32]]
+; SI:  v_cvt_f16_f32_e32 v[[R_F16:[0-9]+]], v[[R_F32]]
+; VI:  v_mov_b32_e32 v[[A_F16:[0-9]+]], 0x3800{{$}}
+; VI:  v_cmp_lt_f16_e32 vcc, v[[A_F16]], v[[B_F16]]
+; VI:  v_cndmask_b32_e32 v[[R_F16:[0-9]+]], v[[D_F16]], v[[C_F16]], vcc
+; GCN: buffer_store_short v[[R_F16]]
+; GCN: s_endpgm
+define void @select_f16_imm_a(
+    half addrspace(1)* %r,
+    half addrspace(1)* %b,
+    half addrspace(1)* %c,
+    half addrspace(1)* %d) {
+entry:
+  %b.val = load half, half addrspace(1)* %b
+  %c.val = load half, half addrspace(1)* %c
+  %d.val = load half, half addrspace(1)* %d
+  %fcmp = fcmp olt half 0xH3800, %b.val
+  %r.val = select i1 %fcmp, half %c.val, half %d.val
+  store half %r.val, half addrspace(1)* %r
+  ret void
+}
+
+; GCN-LABEL: {{^}}select_f16_imm_b
+; GCN: buffer_load_ushort v[[A_F16:[0-9]+]]
+; GCN: buffer_load_ushort v[[C_F16:[0-9]+]]
+; GCN: buffer_load_ushort v[[D_F16:[0-9]+]]
+; SI:  v_cvt_f32_f16_e32 v[[B_F32:[0-9]+]], 0x3800{{$}}
+; SI:  v_cvt_f32_f16_e32 v[[A_F32:[0-9]+]], v[[A_F16]]
+; SI:  v_cmp_lt_f32_e32 vcc, v[[A_F32]], v[[B_F32]]
+; SI:  v_cvt_f32_f16_e32 v[[C_F32:[0-9]+]], v[[C_F16]]
+; SI:  v_cvt_f32_f16_e32 v[[D_F32:[0-9]+]], v[[D_F16]]
+; SI:  v_cndmask_b32_e32 v[[R_F32:[0-9]+]], v[[D_F32]], v[[C_F32]]
+; SI:  v_cvt_f16_f32_e32 v[[R_F16:[0-9]+]], v[[R_F32]]
+; VI:  v_mov_b32_e32 v[[B_F16:[0-9]+]], 0x3800{{$}}
+; VI:  v_cmp_gt_f16_e32 vcc, v[[B_F16]], v[[A_F16]]
+; VI:  v_cndmask_b32_e32 v[[R_F16:[0-9]+]], v[[D_F16]], v[[C_F16]], vcc
+; GCN: buffer_store_short v[[R_F16]]
+; GCN: s_endpgm
+define void @select_f16_imm_b(
+    half addrspace(1)* %r,
+    half addrspace(1)* %a,
+    half addrspace(1)* %c,
+    half addrspace(1)* %d) {
+entry:
+  %a.val = load half, half addrspace(1)* %a
+  %c.val = load half, half addrspace(1)* %c
+  %d.val = load half, half addrspace(1)* %d
+  %fcmp = fcmp olt half %a.val, 0xH3800
+  %r.val = select i1 %fcmp, half %c.val, half %d.val
+  store half %r.val, half addrspace(1)* %r
+  ret void
+}
+
+; GCN-LABEL: {{^}}select_f16_imm_c
+; GCN: buffer_load_ushort v[[A_F16:[0-9]+]]
+; GCN: buffer_load_ushort v[[B_F16:[0-9]+]]
+; GCN: buffer_load_ushort v[[D_F16:[0-9]+]]
+; SI:  v_cvt_f32_f16_e32 v[[C_F32:[0-9]+]], 0x3800{{$}}
+; SI:  v_cvt_f32_f16_e32 v[[A_F32:[0-9]+]], v[[A_F16]]
+; SI:  v_cvt_f32_f16_e32 v[[B_F32:[0-9]+]], v[[B_F16]]
+; SI:  v_cvt_f32_f16_e32 v[[D_F32:[0-9]+]], v[[D_F16]]
+; SI:  v_cmp_lt_f32_e32 vcc, v[[A_F32]], v[[B_F32]]
+; SI:  v_cndmask_b32_e32 v[[R_F32:[0-9]+]], v[[D_F32]], v[[C_F32]]
+; SI:  v_cvt_f16_f32_e32 v[[R_F16:[0-9]+]], v[[R_F32]]
+; VI:  v_cmp_lt_f16_e32 vcc, v[[A_F16]], v[[B_F16]]
+; VI:  v_mov_b32_e32 v[[C_F16:[0-9]+]], 0x3800{{$}}
+; VI:  v_cndmask_b32_e32 v[[R_F16:[0-9]+]], v[[D_F16]], v[[C_F16]], vcc
+; GCN: buffer_store_short v[[R_F16]]
+; GCN: s_endpgm
+define void @select_f16_imm_c(
+    half addrspace(1)* %r,
+    half addrspace(1)* %a,
+    half addrspace(1)* %b,
+    half addrspace(1)* %d) {
+entry:
+  %a.val = load half, half addrspace(1)* %a
+  %b.val = load half, half addrspace(1)* %b
+  %d.val = load half, half addrspace(1)* %d
+  %fcmp = fcmp olt half %a.val, %b.val
+  %r.val = select i1 %fcmp, half 0xH3800, half %d.val
+  store half %r.val, half addrspace(1)* %r
+  ret void
+}
+
+; GCN-LABEL: {{^}}select_f16_imm_d
+; GCN: buffer_load_ushort v[[A_F16:[0-9]+]]
+; GCN: buffer_load_ushort v[[B_F16:[0-9]+]]
+; GCN: buffer_load_ushort v[[C_F16:[0-9]+]]
+; SI:  v_cvt_f32_f16_e32 v[[D_F32:[0-9]+]], 0x3800{{$}}
+; SI:  v_cvt_f32_f16_e32 v[[A_F32:[0-9]+]], v[[A_F16]]
+; SI:  v_cvt_f32_f16_e32 v[[B_F32:[0-9]+]], v[[B_F16]]
+; SI:  v_cvt_f32_f16_e32 v[[C_F32:[0-9]+]], v[[C_F16]]
+; SI:  v_cmp_lt_f32_e32 vcc, v[[A_F32]], v[[B_F32]]
+; SI:  v_cndmask_b32_e32 v[[R_F32:[0-9]+]], v[[D_F32]], v[[C_F32]]
+; SI:  v_cvt_f16_f32_e32 v[[R_F16:[0-9]+]], v[[R_F32]]
+; VI:  v_cmp_lt_f16_e32 vcc, v[[A_F16]], v[[B_F16]]
+; VI:  v_mov_b32_e32 v[[D_F16:[0-9]+]], 0x3800{{$}}
+; VI:  v_cndmask_b32_e32 v[[R_F16:[0-9]+]], v[[D_F16]], v[[C_F16]], vcc
+; GCN: buffer_store_short v[[R_F16]]
+; GCN: s_endpgm
+define void @select_f16_imm_d(
+    half addrspace(1)* %r,
+    half addrspace(1)* %a,
+    half addrspace(1)* %b,
+    half addrspace(1)* %c) {
+entry:
+  %a.val = load half, half addrspace(1)* %a
+  %b.val = load half, half addrspace(1)* %b
+  %c.val = load half, half addrspace(1)* %c
+  %fcmp = fcmp olt half %a.val, %b.val
+  %r.val = select i1 %fcmp, half %c.val, half 0xH3800
+  store half %r.val, half addrspace(1)* %r
+  ret void
+}
+
+; GCN-LABEL: {{^}}select_v2f16
+; SI:  v_cvt_f32_f16_e32
+; SI:  v_cvt_f32_f16_e32
+; SI:  v_cvt_f32_f16_e32
+; SI:  v_cvt_f32_f16_e32
+; SI:  v_cmp_lt_f32_e64
+; SI:  v_cmp_lt_f32_e32
+; VI:  v_cmp_lt_f16_e32
+; VI:  v_cmp_lt_f16_e64
+; GCN: v_cndmask_b32_e32
+; GCN: v_cndmask_b32_e64
+; SI:  v_cvt_f16_f32_e32
+; SI:  v_cvt_f16_f32_e32
+; GCN: s_endpgm
+define void @select_v2f16(
+    <2 x half> addrspace(1)* %r,
+    <2 x half> addrspace(1)* %a,
+    <2 x half> addrspace(1)* %b,
+    <2 x half> addrspace(1)* %c,
+    <2 x half> addrspace(1)* %d) {
+entry:
+  %a.val = load <2 x half>, <2 x half> addrspace(1)* %a
+  %b.val = load <2 x half>, <2 x half> addrspace(1)* %b
+  %c.val = load <2 x half>, <2 x half> addrspace(1)* %c
+  %d.val = load <2 x half>, <2 x half> addrspace(1)* %d
+  %fcmp = fcmp olt <2 x half> %a.val, %b.val
+  %r.val = select <2 x i1> %fcmp, <2 x half> %c.val, <2 x half> %d.val
+  store <2 x half> %r.val, <2 x half> addrspace(1)* %r
+  ret void
+}
+
+; GCN-LABEL: {{^}}select_v2f16_imm_a
+; SI:  v_cvt_f32_f16_e32
+; SI:  v_cvt_f32_f16_e32
+; SI:  v_cvt_f32_f16_e32
+; SI:  v_cvt_f32_f16_e32
+; SI:  v_cmp_gt_f32_e32
+; SI:  v_cvt_f32_f16_e32
+; SI:  v_cvt_f32_f16_e32
+; SI:  v_cvt_f32_f16_e32
+; SI:  v_cvt_f32_f16_e32
+; SI:  v_cmp_gt_f32_e64
+; VI:  v_cmp_lt_f16_e32
+; VI:  v_cmp_lt_f16_e64
+; GCN: v_cndmask_b32_e32
+; SI:  v_cvt_f16_f32_e32
+; GCN: v_cndmask_b32_e64
+; SI:  v_cvt_f16_f32_e32
+; GCN: s_endpgm
+define void @select_v2f16_imm_a(
+    <2 x half> addrspace(1)* %r,
+    <2 x half> addrspace(1)* %b,
+    <2 x half> addrspace(1)* %c,
+    <2 x half> addrspace(1)* %d) {
+entry:
+  %b.val = load <2 x half>, <2 x half> addrspace(1)* %b
+  %c.val = load <2 x half>, <2 x half> addrspace(1)* %c
+  %d.val = load <2 x half>, <2 x half> addrspace(1)* %d
+  %fcmp = fcmp olt <2 x half> <half 0xH3800, half 0xH3900>, %b.val
+  %r.val = select <2 x i1> %fcmp, <2 x half> %c.val, <2 x half> %d.val
+  store <2 x half> %r.val, <2 x half> addrspace(1)* %r
+  ret void
+}
+
+; GCN-LABEL: {{^}}select_v2f16_imm_b
+; SI:  v_cvt_f32_f16_e32
+; SI:  v_cvt_f32_f16_e32
+; SI:  v_cvt_f32_f16_e32
+; SI:  v_cvt_f32_f16_e32
+; SI:  v_cmp_lt_f32_e32
+; SI:  v_cvt_f32_f16_e32
+; SI:  v_cvt_f32_f16_e32
+; SI:  v_cvt_f32_f16_e32
+; SI:  v_cvt_f32_f16_e32
+; SI:  v_cmp_lt_f32_e64
+; VI:  v_cmp_gt_f16_e32
+; VI:  v_cmp_gt_f16_e64
+; GCN: v_cndmask_b32_e32
+; SI:  v_cvt_f16_f32_e32
+; GCN: v_cndmask_b32_e64
+; SI:  v_cvt_f16_f32_e32
+; GCN: s_endpgm
+define void @select_v2f16_imm_b(
+    <2 x half> addrspace(1)* %r,
+    <2 x half> addrspace(1)* %a,
+    <2 x half> addrspace(1)* %c,
+    <2 x half> addrspace(1)* %d) {
+entry:
+  %a.val = load <2 x half>, <2 x half> addrspace(1)* %a
+  %c.val = load <2 x half>, <2 x half> addrspace(1)* %c
+  %d.val = load <2 x half>, <2 x half> addrspace(1)* %d
+  %fcmp = fcmp olt <2 x half> %a.val, <half 0xH3800, half 0xH3900>
+  %r.val = select <2 x i1> %fcmp, <2 x half> %c.val, <2 x half> %d.val
+  store <2 x half> %r.val, <2 x half> addrspace(1)* %r
+  ret void
+}
+
+; GCN-LABEL: {{^}}select_v2f16_imm_c
+; SI:  v_cvt_f32_f16_e32
+; SI:  v_cvt_f32_f16_e32
+; SI:  v_cvt_f32_f16_e32
+; SI:  v_cvt_f32_f16_e32
+; SI:  v_cvt_f32_f16_e32
+; SI:  v_cvt_f32_f16_e32
+; SI:  v_cvt_f32_f16_e32
+; SI:  v_cvt_f32_f16_e32
+; SI:  v_cmp_lt_f32_e32
+; SI:  v_cmp_lt_f32_e64
+; VI:  v_cmp_lt_f16_e32
+; VI:  v_cmp_lt_f16_e64
+; GCN: v_cndmask_b32_e32
+; GCN: v_cndmask_b32_e64
+; SI:  v_cvt_f16_f32_e32
+; SI:  v_cvt_f16_f32_e32
+; GCN: s_endpgm
+define void @select_v2f16_imm_c(
+    <2 x half> addrspace(1)* %r,
+    <2 x half> addrspace(1)* %a,
+    <2 x half> addrspace(1)* %b,
+    <2 x half> addrspace(1)* %d) {
+entry:
+  %a.val = load <2 x half>, <2 x half> addrspace(1)* %a
+  %b.val = load <2 x half>, <2 x half> addrspace(1)* %b
+  %d.val = load <2 x half>, <2 x half> addrspace(1)* %d
+  %fcmp = fcmp olt <2 x half> %a.val, %b.val
+  %r.val = select <2 x i1> %fcmp, <2 x half> <half 0xH3800, half 0xH3900>, <2 x half> %d.val
+  store <2 x half> %r.val, <2 x half> addrspace(1)* %r
+  ret void
+}
+
+; GCN-LABEL: {{^}}select_v2f16_imm_d
+; SI:  v_cvt_f32_f16_e32
+; SI:  v_cvt_f32_f16_e32
+; SI:  v_cvt_f32_f16_e32
+; SI:  v_cvt_f32_f16_e32
+; SI:  v_cvt_f32_f16_e32
+; SI:  v_cvt_f32_f16_e32
+; SI:  v_cvt_f32_f16_e32
+; SI:  v_cvt_f32_f16_e32
+; SI:  v_cmp_lt_f32_e32
+; SI:  v_cmp_lt_f32_e64
+; VI:  v_cmp_lt_f16_e32
+; VI:  v_cmp_lt_f16_e64
+; GCN: v_cndmask_b32_e32
+; GCN: v_cndmask_b32_e64
+; SI:  v_cvt_f16_f32_e32
+; SI:  v_cvt_f16_f32_e32
+; GCN: s_endpgm
+define void @select_v2f16_imm_d(
+    <2 x half> addrspace(1)* %r,
+    <2 x half> addrspace(1)* %a,
+    <2 x half> addrspace(1)* %b,
+    <2 x half> addrspace(1)* %c) {
+entry:
+  %a.val = load <2 x half>, <2 x half> addrspace(1)* %a
+  %b.val = load <2 x half>, <2 x half> addrspace(1)* %b
+  %c.val = load <2 x half>, <2 x half> addrspace(1)* %c
+  %fcmp = fcmp olt <2 x half> %a.val, %b.val
+  %r.val = select <2 x i1> %fcmp, <2 x half> %c.val, <2 x half> <half 0xH3800, half 0xH3900>
+  store <2 x half> %r.val, <2 x half> addrspace(1)* %r
+  ret void
+}